# 5 skills to stop Claude eating your tokens

Claude Code burns context on files, command outputs, and conversation history — and once
the window fills, everything gets slower and pricier. These five moves keep it lean. Three
are built into Claude Code (nothing to install), one is a free open-source plugin, and one
is a custom skill included in this kit.

Honest note up front: the numbers below come from third-party tests and single examples.
They show the *direction* these tools push, not a guarantee. Your results will vary with your
codebase and how you work.

---

## 01 · /doctor — see what your setup costs before it's expensive

`/doctor` is a built-in Claude Code command. Run it and it health-checks your setup and
estimates what your skill list, memory files, and extensions cost in context every session —
then points at the biggest hogs. Run it first so you know where the waste is before you spend
a prompt on it.

- Just type `/doctor` in Claude Code.
- Source: Claude Code skills docs — https://code.claude.com/docs/en/skills

---

## 02 · Ponytail — make Claude write less

Ponytail is a free, MIT-licensed plugin (98k+ stars) that tells the agent to behave like "the
laziest senior dev in the room": write less code and less explanation. Even when total token
counts vary, shorter answers are faster to read and far easier to review.

Install (Claude Code):

    /plugin marketplace add DietrichGebert/ponytail
    /plugin install ponytail@ponytail

What the tests showed (directional, not guaranteed):
- A Reddit experiment measured roughly **-3% to -11% output tokens**.
- An independent JetBrains benchmark on Claude measured about **-15% code** and **-10.3% cost**.

- Repo: https://github.com/DietrichGebert/ponytail
- JetBrains benchmark: https://blog.jetbrains.com/ai/2026/07/ponytail-skill-claude-tested/
- Reddit experiment: https://www.reddit.com/r/ClaudeAI/comments/1v9xjh0/i_tested_5_popular_token_saving_methods_across_10/

---

## 03 · /rewind — undo a wrong path instead of arguing with it

Built into Claude Code. When Claude goes down the wrong path, don't spend ten messages
correcting it (each one adding to context). Run `/rewind` to jump back to the last clean
checkpoint before the mistake and try again.

- Type `/rewind` in Claude Code.
- Source: Claude Code checkpointing docs — https://code.claude.com/docs/en/checkpointing

---

## 04 · /compact — squeeze a bloated session down

Built into Claude Code. `/compact` summarizes the conversation so far and continues from the
summary instead of the full history. One documented community example compressed **167K tokens
into a ~25K summary (about 85%)**. Treat that as a third-party example, not a promise — and
know that compacting can lose detail, so do it at a natural checkpoint.

- Type `/compact` in Claude Code (or let auto-compact handle it).
- Sources: context-window docs — https://code.claude.com/docs/en/context-window ·
  the 85% example — https://camnangai.com/en/for/developers/compact-command-context-rot

---

## 05 · /handoff — carry the work into a fresh session (custom skill, included)

Compacting keeps a session going; sometimes you want a clean `/clear` instead but don't want to
lose the thread. `/handoff` (the custom skill in this kit) writes a compact brief — goal, key
decisions, changed files, next step, and gotchas — that you copy, `/clear`, and paste into the
new session. You start fresh with almost no context but lose nothing that matters.

Install once:
1. Copy the `handoff/` folder from this kit into `~/.claude/skills/` (so you have
   `~/.claude/skills/handoff/SKILL.md`).
2. Restart Claude Code if `/handoff` doesn't show up.

Use it:

    INSTALL ONCE → /handoff → copy the brief → /clear → paste it

- Format reference: Claude Code skills docs — https://code.claude.com/docs/en/skills

---

## The order that works

1. `/doctor` once to see your baseline waste.
2. Install **Ponytail** so answers stay short by default.
3. `/rewind` the moment a path goes wrong — before it compounds.
4. `/compact` when a long session gets heavy but you want to keep going.
5. `/handoff` when you'd rather start clean without losing the plot.

Free and shared by Artem Novitckii — https://novitckii.com